A hybrid model for robust detection of transcription factor binding sites

The short and degenerate nature of transcription factor (TF) binding sites contributes towards a low signal to noise ratio making it very difficult to separate them from their background. In order to tackle this problem one needs to look at ways of capturing the underlying biophysical properties that best discriminates TF binding sites from their background DNA. One such discriminatory property lies in the observed compositional differences in the nucleotide levels of TF binding sites and background DNA which are a result of processes such as purifying selection and selective preferences of TF binding sites for particular nucleotides or a combination of nucleotides over others.

In this article, we present a hybrid model, referred to as a MonoDi-nucleotide model for robustly detecting TF binding sites. It incorporates both mono- and dinucleotide statistics to optimally partition the base positions of an aligned set of TF binding sites (motif) into a non-redundant sequence of mono and/or dinucleotide segments that maximizes the odds ratio of the binding sites relative to their background DNA. We tested the MonoDi-nucleotide model on the benchmark dataset compiled by Tompa et al. (2005) for assessing computational tools that predict TF binding sites. The performance of the MonoDi-nucleotide model on this data set compares well to, and in many cases exceeds, the performance of existing tools. This is in part attributed to the significant role played by dinucleotides in discriminating TF binding sites from background DNA.

A Matlab implementation of the MonoDi-nucleotide model can be found at http://www.utoronto.ca/zhanglab/MonoDi/.